Perplexity has released Decider v1.3 and a smaller Flash v1.3, MarkTechPost reports: open-weight decision models that return probabilities, not paragraphs. You give them a shared text or image input and ask yes/no, multiple-choice or score questions. The main new feature in v1.3 is Locate, which pulls exact text spans out of the input by type or instruction. Weights are under Apache 2.0 on Hugging Face, and both models also run in Perplexity’s Decisions API and on OpenRouter.
The larger model is built on Qwen3.8-27B (Hugging Face lists it around 26B parameters); Flash is a 4B-class distill. MarkTechPost notes you need a CUDA GPU that can hold roughly 49 GiB of BF16 weights for the large model, or about 8.5 GiB for Flash, and there aren’t quantized builds yet. As usual with these announcements, the accuracy numbers are self-reported. Locate is the weakest of the scores they published.
This is the same general idea as Google’s Gemini Decisions API and OpenAI’s Decisions API, only here the weights are actually public.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
Isolate it: try new software in a disposable VM or on a spare machine first, with a snapshot you can roll back to. Don’t point it at a machine you care about.
My take
I like decision models for the boring jobs agents do all day: route this, score that, extract the line that matters. Open weights make that a lot more interesting, because you can run it yourself and test it. Just don’t treat the company’s own leaderboard as the last word, and if you try Locate, check it against text you already know.
God Bless, and Use the Source!
